Agentic Framework Builds Physics-Based World Models via Code

A new paper on arXiv (2605.14398) introduces an agentic framework that constructs physics-based world models through executable simulation code, addressing the physical implausibility of video-based world models. The framework coordinates four agents: planning, code generation, visual review, and physics analysis. The planning agent converts natural language prompts into structured scene plans, the code agent implements them as executable simulation code, and the visual review agent provides visual feedback while the physics analysis agent enforces physical constraints. This approach ensures physically plausible dynamics, avoiding unstable contacts, distorted shapes, or inconsistent motion common in video-based models.